Fish Optimization Of Multi-kernel SVM In The Application Of The Software Defect Prediction

The software industry has become one of the most important indicator to evaluate the comprehensive national strength. With the development of the software industry, software is becoming more and more complex, while the human capacity is still limited, it will inevitably cause a variety of defects, the existence of defects will bring a serious threat to human property, life and health and even national security. Therefore, there is an urgent need for us to address the early detection of potential defects, and software defect prediction technology provides meaningful guidance to solve this problem.One of the purposes of software defect prediction is to judge the tendency of defects in the module through a specific method, that is, to determines whether the module is defective or not. This process can be seen as a pattern recognition process, at the core of it is to classify problems. Therefore, based on the Support Vector Machine Theory, which is widely used in the binary classification, this paper create a defect prediction method. It mainly improves the algorithm from two aspects: the improvement of kernel function and the parameter optimization, and at last, it applies the multi-kernel support vector machine which has been optimized to software defect prediction. The main work includes:(1) Aimed at the problem of Gaussian kernel’s poor generalization effect, I have improved the Gaussian kernel, and by contrast with the traditional Gaussian kernel, improved Gaussian kernel was found not only to maintain the original features of local learning ability, but also improve the classifier generalization ability.(2) Given the powerful capability of the overall kernel’s generalization, and the excellent learning ability of the local kernel function, I put forward a combination of polynomial kernel function and modified Gauss kernel function, and thereby established a hybrid kernel SVM algorithm.(3) Considered the problem that in software defect prediction, the parameters of the classifier are directly affected by the final prediction results, we comparing the several commonly used method of support vector machine parameters selection, and analyzes the influence of classifiers effects and time complexity. It is considered that selecting the AFSA algorithm to improve the effect of the parameters of the support vector machine would lead to a better result, and specific parameter automatic optimization method would be given.(4) I proposed a software defect prediction method which is based on artificial fish swarm optimization multi-kernel support vector machine, and under the NASA MDP data set, i comparison the proposed method with support vector machines, Bayesian method and BP neural network, it turns out that the evaluation index of accuracy, precision and recall of the defect prediction method proposed in this paper are better than others.